dtflow 0.4.3__py3-none-any.whl → 0.5.0__py3-none-any.whl

dtflow/__init__.py

@@ -4,6 +4,7 @@ DataTransformer: 简洁的数据格式转换工具
 核心功能:
 - DataTransformer: 数据加载、转换、保存
 - presets: 预设转换模板 (openai_chat, alpaca, sharegpt, dpo_pair, simple_qa)
+- schema: 数据结构验证 (Schema, Field)
 - tokenizers: Token 统计和过滤
 - converters: HuggingFace/OpenAI 等格式转换
 """
@@ -26,6 +27,23 @@ from .converters import (  # LLaMA-Factory 扩展; ms-swift
 )
 from .core import DataTransformer, DictWrapper, TransformError, TransformErrors
 from .presets import get_preset, list_presets
+from .schema import (
+    Field,
+    Schema,
+    ValidationError,
+    ValidationResult,
+    alpaca_schema,
+    dpo_schema,
+    openai_chat_schema,
+    sharegpt_schema,
+    validate_data,
+)
+from .framework import (
+    CompatibilityResult,
+    check_compatibility,
+    detect_format,
+    export_for,
+)
 from .storage import load_data, sample_file, save_data
 from .streaming import StreamingTransformer, load_sharded, load_stream, process_shards
 from .tokenizers import (
@@ -42,7 +60,7 @@ from .tokenizers import (
     token_stats,
 )
 
-__version__ = "0.4.3"
+__version__ = "0.5.0"
 
 __all__ = [
     # core
@@ -53,6 +71,21 @@ __all__ = [
     # presets
     "get_preset",
     "list_presets",
+    # schema
+    "Schema",
+    "Field",
+    "ValidationResult",
+    "ValidationError",
+    "validate_data",
+    "openai_chat_schema",
+    "alpaca_schema",
+    "dpo_schema",
+    "sharegpt_schema",
+    # framework
+    "CompatibilityResult",
+    "check_compatibility",
+    "detect_format",
+    "export_for",
     # storage
     "save_data",
     "load_data",

dtflow/__main__.py

@@ -18,6 +18,7 @@ Commands:
     clean        数据清洗
     run          执行 Pipeline 配置文件
     history      显示数据血缘历史
+    validate     使用 Schema 验证数据格式
     mcp          MCP 服务管理（install/uninstall/status）
     logs         日志查看工具使用说明
 """
@@ -40,6 +41,7 @@ from .cli.commands import stats as _stats
 from .cli.commands import tail as _tail
 from .cli.commands import token_stats as _token_stats
 from .cli.commands import transform as _transform
+from .cli.commands import validate as _validate
 
 # 创建主应用
 app = typer.Typer(
@@ -211,6 +213,26 @@ def history(
     _history(filename, json)
 
 
+# ============ 验证命令 ============
+
+
+@app.command()
+def validate(
+    filename: str = typer.Argument(..., help="输入文件路径"),
+    preset: Optional[str] = typer.Option(
+        None, "--preset", "-p", help="预设 Schema: openai_chat, alpaca, dpo, sharegpt"
+    ),
+    output: Optional[str] = typer.Option(None, "--output", "-o", help="输出有效数据的文件路径"),
+    filter: bool = typer.Option(
+        False, "--filter", "-f", help="过滤无效数据并保存"
+    ),
+    max_errors: int = typer.Option(20, "--max-errors", help="最多显示的错误数量"),
+    verbose: bool = typer.Option(False, "--verbose", "-v", help="显示详细信息"),
+):
+    """使用预设 Schema 验证数据格式"""
+    _validate(filename, preset, output, filter, max_errors, verbose)
+
+
 # ============ 工具命令 ============
 
 

dtflow/cli/commands.py

@@ -33,6 +33,9 @@ from .pipeline import run
 # 血缘追踪命令
 from .lineage import history
 
+# 验证命令
+from .validate import validate
+
 __all__ = [
     # 采样
     "sample",
@@ -53,4 +56,6 @@ __all__ = [
     "run",
     # 血缘
     "history",
+    # 验证
+    "validate",
 ]

dtflow/cli/validate.py

@@ -0,0 +1,152 @@
+"""
+CLI Schema 验证命令
+"""
+
+from pathlib import Path
+from typing import Optional
+
+from ..schema import (
+    Schema,
+    Field,
+    alpaca_schema,
+    dpo_schema,
+    openai_chat_schema,
+    sharegpt_schema,
+)
+from ..storage.io import load_data, save_data
+from .common import _check_file_format
+
+
+# 预设 Schema 映射
+PRESET_SCHEMAS = {
+    "openai_chat": openai_chat_schema,
+    "openai-chat": openai_chat_schema,
+    "chat": openai_chat_schema,
+    "alpaca": alpaca_schema,
+    "dpo": dpo_schema,
+    "dpo_pair": dpo_schema,
+    "sharegpt": sharegpt_schema,
+}
+
+
+def validate(
+    filename: str,
+    preset: Optional[str] = None,
+    output: Optional[str] = None,
+    filter_invalid: bool = False,
+    max_errors: int = 20,
+    verbose: bool = False,
+) -> None:
+    """
+    使用 Schema 验证数据文件。
+
+    Args:
+        filename: 输入文件路径
+        preset: 预设 Schema 名称 (openai_chat, alpaca, dpo, sharegpt)
+        output: 输出文件路径（保存有效数据）
+        filter_invalid: 过滤无效数据并保存
+        max_errors: 最多显示的错误数量
+        verbose: 显示详细信息
+
+    Examples:
+        dt validate data.jsonl --preset=openai_chat
+        dt validate data.jsonl --preset=alpaca -o valid.jsonl
+        dt validate data.jsonl --preset=chat --filter
+    """
+    filepath = Path(filename)
+
+    if not filepath.exists():
+        print(f"错误: 文件不存在 - {filename}")
+        return
+
+    if not _check_file_format(filepath):
+        return
+
+    # 确定 Schema
+    if preset is None:
+        # 列出可用的预设
+        print("请指定预设 Schema (--preset):")
+        print()
+        for name in ["openai_chat", "alpaca", "dpo", "sharegpt"]:
+            print(f"  --preset={name}")
+        print()
+        print("示例:")
+        print(f"  dt validate {filename} --preset=openai_chat")
+        return
+
+    preset_lower = preset.lower().replace("-", "_")
+    if preset_lower not in PRESET_SCHEMAS:
+        print(f"错误: 未知的预设 Schema '{preset}'")
+        print(f"可用预设: {', '.join(['openai_chat', 'alpaca', 'dpo', 'sharegpt'])}")
+        return
+
+    schema = PRESET_SCHEMAS[preset_lower]()
+
+    # 加载数据
+    try:
+        data = load_data(str(filepath))
+    except Exception as e:
+        print(f"错误: 无法读取文件 - {e}")
+        return
+
+    if not data:
+        print("文件为空")
+        return
+
+    total = len(data)
+    print(f"验证文件: {filepath.name}")
+    print(f"预设 Schema: {preset}")
+    print(f"总记录数: {total}")
+    print()
+
+    # 验证
+    valid_data = []
+    invalid_count = 0
+    error_samples = []
+
+    for i, item in enumerate(data):
+        result = schema.validate(item)
+        if result.valid:
+            valid_data.append(item)
+        else:
+            invalid_count += 1
+            if len(error_samples) < max_errors:
+                error_samples.append((i, result))
+
+    valid_count = len(valid_data)
+    valid_ratio = valid_count / total * 100 if total > 0 else 0
+
+    # 输出结果
+    if invalid_count == 0:
+        print(f"✅ 全部通过! {valid_count}/{total} 条记录有效 (100%)")
+    else:
+        print(f"⚠️ 验证结果: {valid_count}/{total} 条有效 ({valid_ratio:.1f}%)")
+        print(f"   无效记录: {invalid_count} 条")
+        print()
+
+        # 显示错误示例
+        print(f"错误示例 (最多显示 {max_errors} 条):")
+        print("-" * 60)
+
+        for idx, result in error_samples:
+            print(f"[第 {idx} 行]")
+            for err in result.errors[:3]:  # 每条记录最多显示 3 个错误
+                print(f"  - {err}")
+            if len(result.errors) > 3:
+                print(f"  ... 还有 {len(result.errors) - 3} 个错误")
+            print()
+
+    # 保存有效数据
+    if output or filter_invalid:
+        output_path = output or str(filepath).replace(
+            filepath.suffix, f"_valid{filepath.suffix}"
+        )
+        save_data(valid_data, output_path)
+        print(f"✅ 有效数据已保存: {output_path} ({valid_count} 条)")
+
+    # 详细模式：显示 Schema 定义
+    if verbose:
+        print()
+        print("Schema 定义:")
+        print("-" * 40)
+        print(schema)

dtflow/core.py

@@ -386,6 +386,88 @@ class DataTransformer:
 
         return errors
 
+    def validate_schema(
+        self,
+        schema: "Schema",
+        on_error: Literal["skip", "raise", "filter"] = "skip",
+        max_errors: int = 100,
+    ) -> Union["DataTransformer", List[tuple]]:
+        """
+        使用 Schema 验证数据结构。
+
+        Args:
+            schema: Schema 对象，定义数据结构验证规则
+            on_error: 错误处理方式
+                - "skip": 打印警告，返回验证失败的记录列表
+                - "raise": 第一个错误时抛出异常
+                - "filter": 过滤掉验证失败的记录，返回新的 DataTransformer
+            max_errors: 最大错误数量（on_error="skip" 时生效）
+
+        Returns:
+            - on_error="skip": 返回 [(index, ValidationResult), ...] 失败记录列表
+            - on_error="raise": 无返回（成功）或抛出 ValueError
+            - on_error="filter": 返回过滤后的新 DataTransformer
+
+        Examples:
+            >>> from dtflow import Schema, Field
+            >>> schema = Schema({
+            ...     "messages": Field(type="list", required=True, min_length=1),
+            ...     "messages[*].role": Field(type="str", choices=["user", "assistant"]),
+            ... })
+
+            >>> # 获取验证失败的记录
+            >>> errors = dt.validate_schema(schema)
+            >>> for idx, result in errors:
+            ...     print(f"第 {idx} 行验证失败: {result.errors}")
+
+            >>> # 过滤掉无效记录
+            >>> valid_dt = dt.validate_schema(schema, on_error="filter")
+
+            >>> # 遇到错误立即停止
+            >>> dt.validate_schema(schema, on_error="raise")
+        """
+        from .schema import Schema, ValidationResult
+
+        failed: List[tuple] = []
+        valid_data: List[dict] = []
+        error_count = 0
+
+        for i, item in enumerate(self._data):
+            result = schema.validate(item)
+            if result.valid:
+                valid_data.append(item)
+            else:
+                failed.append((i, result))
+                error_count += len(result.errors)
+
+                if on_error == "raise":
+                    error_msgs = [str(e) for e in result.errors[:3]]
+                    raise ValueError(
+                        f"第 {i} 行验证失败:\n  " + "\n  ".join(error_msgs)
+                    )
+
+                if on_error == "skip" and error_count >= max_errors:
+                    print(f"⚠️ 已达到最大错误数 {max_errors}，停止验证")
+                    break
+
+        if on_error == "skip":
+            if failed:
+                print(f"⚠️ 验证失败 {len(failed)} 条记录（共 {error_count} 个错误）")
+            return failed
+
+        if on_error == "filter":
+            tracker = self._lineage_tracker
+            if tracker:
+                tracker.record(
+                    "validate_schema",
+                    {"schema": repr(schema), "on_error": on_error},
+                    len(self._data),
+                    len(valid_data),
+                )
+            return DataTransformer(valid_data, _lineage_tracker=tracker)
+
+        return failed
+
     def dedupe(
         self,
         key: Union[None, str, List[str], Callable[[Any], Any]] = None,
@@ -801,6 +883,78 @@ class DataTransformer:
         filtered = [item for item, keep in zip(self._data, mask) if keep]
         return DataTransformer(filtered)
 
+    # ============ 训练框架集成 ============
+
+    def check_compatibility(
+        self,
+        framework: Literal["llama-factory", "swift", "axolotl"],
+    ) -> "CompatibilityResult":
+        """
+        检查数据与目标训练框架的兼容性。
+
+        Args:
+            framework: 目标框架名称
+                - "llama-factory": LLaMA-Factory
+                - "swift": ms-swift (ModelScope)
+                - "axolotl": Axolotl
+
+        Returns:
+            CompatibilityResult 对象，包含 valid, errors, warnings, suggestions
+
+        Examples:
+            >>> result = dt.check_compatibility("llama-factory")
+            >>> if result.valid:
+            ...     print("兼容!")
+            >>> else:
+            ...     print(result.errors)
+        """
+        from .framework import check_compatibility
+
+        return check_compatibility(self._data, framework)
+
+    def export_for(
+        self,
+        framework: Literal["llama-factory", "swift", "axolotl"],
+        output_dir: str,
+        dataset_name: str = "custom_dataset",
+        **kwargs,
+    ) -> Dict[str, str]:
+        """
+        一键导出数据和配置文件到目标训练框架。
+
+        Args:
+            framework: 目标框架名称
+            output_dir: 输出目录
+            dataset_name: 数据集名称
+            **kwargs: 框架特定参数（如 model_name）
+
+        Returns:
+            生成的文件路径字典 {"data": "...", "config": "...", ...}
+
+        Examples:
+            >>> # 导出到 LLaMA-Factory
+            >>> dt.export_for("llama-factory", "./llama_ready")
+            # 生成:
+            # - ./llama_ready/custom_dataset.json
+            # - ./llama_ready/dataset_info.json
+            # - ./llama_ready/train_args.yaml
+
+            >>> # 导出到 ms-swift
+            >>> dt.export_for("swift", "./swift_ready", dataset_name="my_data")
+
+            >>> # 导出到 Axolotl
+            >>> dt.export_for("axolotl", "./axolotl_ready")
+        """
+        from .framework import export_for
+
+        return export_for(
+            self._data,
+            framework,
+            output_dir,
+            dataset_name=dataset_name,
+            **kwargs,
+        )
+
 
 def _sanitize_key(name: str) -> str:
     """将字段名规范化为合法的 Python 标识符"""